Update to simde v0.8.2

.codespellrc

@@ -1,3 +1,3 @@
 [codespell]
-skip = ./third_party/simde,./.git,./benchmark/python,./experiments,./gh-pages,./build
+skip = ./third_party/simde-0.8.2,./.git,./benchmark/python,./experiments,./gh-pages,./build
 ignore-words-list=shft

CMakeLists.txt

@@ -159,7 +159,7 @@ install (
   DESTINATION include
   FILES_MATCHING PATTERN "*.hpp")
 install (
-  DIRECTORY ${CMAKE_SOURCE_DIR}/third_party/simde
+  DIRECTORY ${CMAKE_SOURCE_DIR}/third_party/simde-0.8.2
   DESTINATION include/${HPCOMBI_INSTALL_DIR})
 install (
   FILES ${CMAKE_CURRENT_BINARY_DIR}/hpcombi.pc

examples/Renner.cpp

@@ -31,9 +31,8 @@
 #include <unordered_map>
 #endif
 
-#include "simde/x86/sse4.1.h"  // for simde_mm_max_epu8, simde...
 #include "hpcombi/perm16.hpp"
-
+#include "simde-0.8.2/x86/sse4.1.h"  // for simde_mm_max_epu8, simde...
 
 template <typename T>
 std::ostream &operator<<(std::ostream &out, const std::vector<T> &v) {

examples/pattern.cpp

@@ -30,21 +30,21 @@
 #else
 #include <unordered_map>
 #endif
-#include "simde/x86/sse4.1.h"  // for simde_mm_max_epu8, simde...
+#include "simde-0.8.2/x86/sse4.1.h"  // for simde_mm_max_epu8, simde...
 
 #include "hpcombi/perm16.hpp"
 
 template <typename T>
 std::ostream &operator<<(std::ostream &out, const std::vector<T> &v) {
-  out << '[';
-  if (!v.empty()) {
-    auto i = v.begin();
-    for (; i != --v.end(); ++i)
-        out << std::setw(2) << *i << ",";
-    out << std::setw(2) << *i;
-  }
-  out << "]";
-  return out;
+    out << '[';
+    if (!v.empty()) {
+        auto i = v.begin();
+        for (; i != --v.end(); ++i)
+            out << std::setw(2) << *i << ",";
+        out << std::setw(2) << *i;
+    }
+    out << "]";
+    return out;
 }
 
 using namespace std;
@@ -54,16 +54,18 @@ std::vector<epu8> subsets;
 std::vector<epu8> subperm;
 
 epu8 tosubset(uint16_t n) {
-    epu8 res {};
+    epu8 res{};
     for (int i = 0; i < 16; i++) {
-        if (((n >> i) & 1) != 0) res[i] = 0xff;
+        if (((n >> i) & 1) != 0)
+            res[i] = 0xff;
     }
-    if (simde_mm_movemask_epi8(res) != n) cout << n << "BUG" << res << endl;
+    if (simde_mm_movemask_epi8(res) != n)
+        cout << n << "BUG" << res << endl;
     return res;
 }
 
 epu8 subset_to_perm(epu8 s) {
-    epu8 res = Epu8({},0xff);
+    epu8 res = Epu8({}, 0xff);
     int c = 0;
     for (int i = 0; i < 16; i++) {
         if (s[i] != 0) {
@@ -76,27 +78,26 @@ epu8 subset_to_perm(epu8 s) {
 
 void make_subsets_of_size(int n, int k) {
     int n2 = 1 << n;
-    for (uint16_t i=0; i < n2; i++) {
+    for (uint16_t i = 0; i < n2; i++) {
         if (__builtin_popcountl(i) == k) {
             subsets.push_back(tosubset(i));
             subperm.push_back(subset_to_perm(tosubset(i)));
         }
     }
 }
 
-template <int Size>
-epu8 extract_pattern(epu8 perm, epu8 permset) {
+template <int Size> epu8 extract_pattern(epu8 perm, epu8 permset) {
     epu8 cst = Epu8({}, Size);
     epu8 res = permuted(perm, permset) | (Epu8.id() >= cst);
     res = sort_perm(res) & (Epu8.id() < cst);
     return res;
 }
 
-template <int Size>
-bool has_pattern(epu8 perm, epu8 patt) {
+template <int Size> bool has_pattern(epu8 perm, epu8 patt) {
     for (size_t i = 0; i < subperm.size(); i++) {
         epu8 extr = extract_pattern<Size>(perm, subperm[i]);
-        if (equal(extr, patt)) return true;
+        if (equal(extr, patt))
+            return true;
     }
     return false;
 }
@@ -106,13 +107,13 @@ int main() {
     int n = 8, k = 4, n2 = 1 << n;
     make_subsets_of_size(n, k);
     cout << subsets.size() << endl;
-    epu8 perm = {1,4,2,0,3,5,6,7};
+    epu8 perm = {1, 4, 2, 0, 3, 5, 6, 7};
     int i = 42;
     cout << Perm16::one() << endl;
     cout << perm << endl;
     cout << subsets[i] << endl;
     cout << simde_mm_movemask_epi8(subsets[i]) << endl;
     cout << extract_pattern<4>(perm, subperm[i]) << endl;
-    cout << int(has_pattern<4>(perm, epu8 {2,1,0,3})) << endl;
-    cout << int(has_pattern<4>(perm, epu8 {3,2,1,0})) << endl;
+    cout << int(has_pattern<4>(perm, epu8{2, 1, 0, 3})) << endl;
+    cout << int(has_pattern<4>(perm, epu8{3, 2, 1, 0})) << endl;
 }

include/hpcombi/bmat16.hpp

@@ -34,7 +34,7 @@
 #include "bmat8.hpp"  // for BMat8
 #include "debug.hpp"  // for HPCOMBI_ASSERT
 
-#include "simde/x86/avx2.h"
+#include "simde-0.8.2/x86/avx2.h"
 
 namespace HPCombi {
 using xpu16 = uint16_t __attribute__((vector_size(32)));

include/hpcombi/epu8.hpp

@@ -45,8 +45,8 @@ eg simde_mm_testz_si128(a,a) → is_all_zero(a) */
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wbitwise-instead-of-logical"
 #endif
-#include "simde/x86/sse4.1.h"  // for simde_mm_max_epu8, simde...
-#include "simde/x86/sse4.2.h"  // for ???
+#include "simde-0.8.2/x86/sse4.1.h"  // for simde_mm_max_epu8, simde...
+#include "simde-0.8.2/x86/sse4.2.h"  // for ???
 #if defined(__clang__)
 #pragma clang diagnostic pop
 #endif

include/hpcombi/perm16.hpp

@@ -48,8 +48,8 @@
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wbitwise-instead-of-logical"
 #endif
-#include "simde/x86/sse4.1.h"
-#include "simde/x86/sse4.2.h"
+#include "simde-0.8.2/x86/sse4.1.h"
+#include "simde-0.8.2/x86/sse4.2.h"
 #if defined(__clang__)
 #pragma clang diagnostic pop
 #endif